Parameter Set for Computer-Assisted Texture Analysis of Fetal Brain

Gentillon et al. BMC Res Notes (2016) 9:496 DOI 10.1186/s13104-016-2300-3 BMC Research Notes

RESEARCH ARTICLE Open Access Parameter set for computer‑assisted texture analysis of fetal brain Hugues Gentillon1,2*, Ludomir Stefańczyk1, Michał Strzelecki2 and Maria Respondek‑Liberska3

Abstract Background: Magnetic resonance data were collected from a diverse population of gravid women to objectively compare the quality of 1.5-tesla (1.5 T) versus 3-T magnetic resonance imaging of the developing human brain. MaZda and B11 computational-visual cognition tools were used to process 2D images. We proposed a wavelet-based parameter and two novel histogram-based parameters for Fisher texture analysis in three-dimensional space. Results: Wavenhl, focus index, and dispersion index revealed better quality for 3 T. Though both 1.5 and 3 T images were 16-bit DICOM encoded, nearly 16 and 12 usable bits were measured in 3 and 1.5 T images, respectively. The four- bit padding observed in 1.5 T K-space encoding mimics noise by adding illusionistic details, which are not really part of the image. In contrast, zero-bit padding in 3 T provides space for storing more details and increases the likelihood of noise but as well as edges, which in turn are very crucial for differentiation of closely related anatomical structures. Conclusions: Both encoding modes are possible with both units, but higher 3 T resolution is the main difference. It contributes to higher perceived and available dynamic range. Apart from surprisingly larger Fisher coefficient, no significant difference was observed when testing was conducted with down-converted 8-bit BMP images. Keywords: Histogram, Wavelets, Computer-assisted radiology, Hugues Gentillon, Teleradiology, Prenatal development, Fetal brain, Mazda, b11, Medical cybernetics, Artificial intelligence, Computational visual cognition

Background MRI replacing USG during pregnancy examination Computer is a non-biological copy of the human brain. In terms of safety, ultrasonography (USG) remains the As its use increases day-by-day in medicine, various gold standard for prenatal central nervous system (CNS) transdisciplinary approaches emerge. Among them is imaging [1]. It is used for prevention and diagnosis of texture analysis, the evolving cybernetics of radiology. congenital malformation [2, 3]. Nevertheless, there are This experiment is a translational study seeking to objec- cases which benefit from alternative imaging techniques tively compare the quality of 1.5-tesla (1.5 T) versus 3-T like 1.5 or 3 T MRI [4–9]. Besides spotting maternal magnetic resonance imaging (MRI) of the developing abnormalities, supplemental information from magnetic human brain (Fig. 1), in order to determine whether the resonance imaging is also crucial for identification of extra administrative cost is worthy for the patient and fetal anatomy and pathology [10, 11]. MRI is also being the healthcare system. First there was a need to develop increasingly used as correlative imaging modality in an objective methodology to collect and process the data pregnancy—because it uses no ionizing radiation, has and subsequently justify its necessity and dispel distrust no known teratogenic effects, provides excellent soft-tis- of financial burden. sue contrast, and has multiple planes for reconstruction and large field of view, allowing better depiction of neu- roanatomy in fetuses with large or complex anomalies [12–15]. Compared to magnetic resonance imaging, USG is more affordable and widely used in many countries, *Correspondence: [email protected] 1 Department of Radiology and Diagnostic Imaging, Barlicki University as a radiologic tool in routine examination of pregnant Hospital, Medical University of Lodz, Lodz, Poland women, especially in the detection of fetal anomalies, at Full list of author information is available at the end of the article about 20th week of gestation [2, 3]. On the other hand,

© The Author(s) 2016. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/ publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Gentillon et al. BMC Res Notes (2016) 9:496 Page 2 of 18

Fig. 2 Ultrasonography (USG) vs. magnetic resonance imaging (MRI). Fig. 1 “Do you see what I see?”. Fetal brain. a Left 1.5 T. b Right 3 T. This Fetal MRI was used as a complimentary modality to USG. a USG. b figure was incorporated and so-titled in this manuscript to illustrate MRI that magnetic resonance (MR) images can be used for evaluation of achromatic vision and sensitivity to change in grayscale quality between different subjects. For fair comparison, this exercise should be blinded. It is recommended to have at least two radiologists, for integration into medical practice. To this date, texture if not available, medically-trained practitioners or any volunteers analysis is used in pre-diagnosis of the globally pandemic (blinded) and one examiner (also blinded) to look at the images side-by-side, on two medical-diagnostic monitors, engineered for disease of tuberculosis [27–29]. In Bangladesh, for exam- 16-bit display (same brand/model in natively flat display mode: i.e. ple, radiologists and radio-technicians quickly screen without any added enhancements). All in-computer/in-monitor RT patients displaying tuberculosis (TB) signs with CAD4TB (real-time) editing features must be turned off (incl. hardware/soft‑ texture analysis software [30–33]. This medical applica- ware rendering): n.b. 16-bit of true data has more room for “contrast tion of texture analysis is sponsored by The World Health booster”—which is essentially an illusion, as a result of post-editing artifacts, not really part of the image. In this investigation, volunteers Organization, at a mere cost of $3 per CAD4TB test. were asked to identify the images unlabeled (, of course). It was Only patients pre-diagnosed TB-positive by CAD4TB observed that both images were equally sharp (in pass-through software receive the more expensive molecular examina- mode), for a normal, naked human eye. There flows the explanation tion known as GeneXpert TB Diagnosis and Resistance for the research goal to mathematically determine which MR modal‑ Test [30–33]. In the aforementioned example, an increase ity actually produces images with more captured details. Both images were 16-bit encoded, and the difference could not be measured via in the rate of early- detection of TB has been reported in “perceived dynamic range” (visible details). With computer vision clinics where texture analysis is used as a pre-diagnostic software, 1.5/3 T images can be numerically decoded to accurately tool [30–33]. Apart from image encoding statistics, other assess “available dynamic range” (visible/invisible details) factors also affect how clinicians perceive medical image appearance: e.g. acutance or subjective quality factor, human eye’s contrast sensitivity function, modulation 1.5 and 3 T MRI scanners produce superior CNS images transfer function or spatial frequency response, image (Fig. 2) but are not cost-effectively built. These units are displayed height, viewing distance, positioning of the massive and expensive [16–20]. They are mostly availa- subjects (stand, seat), radiology LCD (liquid crystal dis- ble in big-budget hospitals and wealthy medical research play) screen characteristics (anti-glare, glossy, size, reso- centers. Because of such disadvantages, carrying research lution, calibration), etc [34–36]. with MRI is more likely to be impeded. As a result, little information is known about its relevance to potential Wavelets‑based texture analysis beneficiaries. MRI is bureaucratically recommended only Wavelets are texture analysis parameters which classify if it is essential in some health care systems administrated intrinsic properties of an image into matrix layers, and by a central office (e.g. Sweden) [21]. Physicians are even therefore they are very powerful tools for differentiating liable for part of MRI examination cost in some territo- and matching patterns in regions of interest. Wavelets are ries (e.g. Poland) [22, 23]. Thus relevance to beneficiaries used, for example, in forensic investigation to match fin- is likely to be affected in such health care systems. gerprints [37], in diagnostic radiology to detect suspected diseases [38], and in many other applications where digi- What is texture analysis? tal imaging is needed [37]. Texture analysis is an artificial process involving quantification of image quality by means of parametric features Histogram‑based texture analysis to characterize regions of interest (ROI) [24–26]. While To fully grasp the concept behind the methodology used texture analysis is still far from replacing the clinician in this experiment, it is important to first comprehend or eye, some of its features warrant further consideration at least have a minimum knowledge about the purpose, Gentillon et al. BMC Res Notes (2016) 9:496 Page 3 of 18

functionality, and usefulness of histogram in digital imag- bits = 28 = 256). In other words, a 16-bit image would be ing. Just as the aforementioned brain images in Fig. 1, the down-sample in order to display its geometric represen- quality and texture of every shape in Fig. 3 may not be tation on such computational histograms (Figs. 4, 5, 6). A interpreted in the same way by different pairs of human Histogram display in MaZda, on the other hand, does not eyes. graphically display pixel values larger than 255. However, Histogram parameters are digital imaging tools which it is able to quantify 16-bit DICOM (Digital Imaging and analyze textural characteristics of a picture by measuring Communications in Medicine), evaluate quality of differ- their quality [39, 40]. Such parameters can differentiate ent MR modalities, and differentiate ROIs. Is this limit a texture by quantifying, for example, resolution, sharp- histogramatic error? No, it is not. A radiology monitor of, ness, noise, actual edge details and artifacts, negative probably, the size of a theatre screen is needed to visually and positive exposures, as well as aliasing and moiré— display a linear representation of every value in a 16-bit 16 i.e. superimposed patterns in MR images [41–44]; moiré image (2 − 1 = 65,535). In spite of the graphical limita- effect is most commonly observed in uncontrolled tions, histogram-based texture analysis is more efficient TV interviews recorded with older cameras—and the than edge detection for pixel segmentation and identi- interviewees were wearing pinstriped clothing [45, 46]. fication of region boundaries [47, 48]. Some histogram In some software, histogram is customized for 8-bit parameters (e.g. mean, variance, kurtosis, skewness) can graphical visualization to fit the display of conven- be used to differentiate noise and related artifacts from tional RGB monitors, with luminance value of 0–255 (8 very sharp details such as edges (region boundaries).

Fig. 3 Varying shades of grey. a Arrange from lighter to darker? b Answers to Puzzle as per measurements from computer vision software. Again, test subjects should be blinded

Fig. 4 Histogram representation of fetal brain. a MRI of fetal brain. b Histogram of the whole image Gentillon et al. BMC Res Notes (2016) 9:496 Page 4 of 18

Fig. 5 Regions of interest. a ROI 1 ventricles; 2 thalamus; 3 grey matter; 4 white matter. b Histogram parameters from texture analysis with MaZda

Fig. 6 ROI analysis. a ROI 1 Ventricles; b ROI 2 Thalamic nuclei; c ROI 3 Grey matter; d ROI 4 White matter. There exist several techniques and methods of texture analysis. Non-parametric graphs (e.g. histogram, box plot) would be indeed a simple alternative to conduct this study. As shown in the figures, MaZda does display histogram of 8-bit but not for 16-bit DICOM image. Also, there are issues with drawing conclusions straight from histograms and boxplots. They are pictorial representations and thus are indirect methods. Furthermore non-parametric interpretation may not be as precise and accurate as parametric quantification. Therefore, parametric quantification was used to assess image quality rather than conventional appearance. 3D, non-parametric graphs are also possible with collateral usage of MaZda (version 5) and B11 (version 3.3) in training mode. The problem with training methods is that errors might occur as a result of overtraining the network. Hence, raw analysis was performed

Methods of Lodz (MUL), Barlicki University Hospitals (BUH), This multidisciplinary research was approved by a rele- Polish Mother’s Memorial Hospital Research Institute vant bioethics committee, consisting of emeritus profes- (ICZMP) et al. Written informed consent was obtained sors and senior researchers from the Medical University from all subjects, and the methods were carried out in Gentillon et al. BMC Res Notes (2016) 9:496 Page 5 of 18

accordance with the approved guidelines (see end of the 70 years ago [50]. This background history explained why manuscript for registration number). The experiment patients travelled there to seek second opinion rather was divided into three parts. than just for the known association of cost-effective treatment in Poland. The selected samples were diverse— Research tools i.e. subjects were not closely related (not monozygotic Firstly, some quality-control checks were performed twins, not dizygotic twins, not consanguineous twins, before designing this research. Images (Fig. 1) were etc). Gestational ages were between 20 and 40 weeks. imported and uncompressed in Photoshop CS6 64-bit Patients underwent either 1.5 T scan at BUH or 3 T scan Extended. The software was set to HDR (high-bit-depth at ICZMP—but not both. It was deemed irrelevant in resolution) mode (i.e. floating-point numeric representa- terms of assessment of image quality as well as for good tion: up to 32 bits × 3 channels). RAW-formatted images patient care and fetal rights (i.e. moral and legal rights were simultaneously examined with radiology computer of human fetuses) [51]. In recent years, physical appear- systems (i.e. high-end hardware/accessories, monitors ances of subjects (such as eye shapes) have been docu- capable to render 16-bit grayscale natively; memory- mented to cause some photography cameras to capture intensive controller/graphic card, Phillips DICOM deceptive images in auto-mode—a dysfunction due to Viewer R3.0 SP3, and LCD calibration software). With faulty-engineering and bug compatibility [52–54]. To our proper-matching adjustments, difference between the knowledge, such phenomena have not been reported in two images were imperceptibly unnoticeable. That was a radiological imaging. Furthermore it is the responsibil- conundrum which proved the need to carry out quality- ity of manufacturers to ensure that MR units are capable assessment research with MaZda and B11. to capture good quality sequences, regardless of mother and/or fetus phenotypes. In this investigation, the arti- Trials facts observed in excluded samples were consequences BUH (1.5 T unit) and ICZMP (3 T unit) radiology depart- of fetal movement and inadequate settings by radiog- ments prescribed MRI to patients for further differential- raphers, not due to phenotypic characteristics. Beside diagnostic investigation—rather than just a concomitant consent-and-publication agreement, the objective of this adjunct to ultrasound. 288 MR images of normal fetal research did not rationally necessitate broadcasting fur- brain were used in the main assessment and 72 MR ther descriptions about patients’ background/identities images of normal fetal brain in the supplementary trials (incl. phenotypes and anamneses) [51]. Therefore, such (see Table 1 for further description). details were omitted in this manuscript as well as in the MR images came from MRI studies of different patients datasets [51]. who had fetal MRI as a complimentary modality to ultra- MaZda software package 5 (B11 included) was used for sonography and echocardiography, to further assess and quantification of MR images. A wavelet-based parameter confirm the diagnosis of suspected anomalies affect- (wavenhl) was combined with two novel histogram-based ing the fetuses as well as the pregnant patients. Most parameters (focus index, dispersion index) to perform participants were seeking remedies for conditions not Fisher-texture analysis in three-dimensional space. It was related to fetal brain. Research hospitals in Poland are hypothesized that Wavenhl could fingerprint (match) among the forefront pioneers in prenatal diagnosis of regions of interest—i.e. as per human anatomy, the build- congenital malformation, in utero surgery, cardiology ing blocks of normal thalamus, for example, is the same and minimally-invasive cardio-surgery [49]. In spite of regardless of subjects and MR modalities. Two param- its stigmatic plight, modern cardiology arguably began in eters which were not sensitive to minute variations in Lodz, Poland, with advances in medical electronics about phenotypes were favored in this research. Hence, focus index and dispersion index were utilized for assessing image quality in terms of resolution, sharpness, aliasing Table 1 Sample description and moiré, noise, actual edge details and artifacts. Mic- 1.5 T 3 T Total roDicom was used to extract images from teleradiology network systems and storage media containing raw data A. Main trials 144 144 288 of 3 and 1.5 T MRI studies. One DICOM and one BMP B. Supplementary trials 36 36 72 (also known as Bitmap: Microsoft Windows Device- Group 1 20 20 40 Independent Bitmap file) were generated from each MR Group 2 16 16 32 sequence to create four batches of images: 3 T DICOM, A. Main trials 288 MR images, categorized by magnetic field strength of scanner 1.5 T DICOM, 3 T BMP, and 1.5 T BMP. All files were (144 1.5 and 144 3 T). B. Supplementary trials 72 MR images (36 1.5 and × × × 36 3 T). Samples used in the supplementary trials were further divided into extracted in their native size and resolution. Both 3 × two categories, by gestational age: group 1 20–28 weeks; group 2 29–40 weeks and 1.5 T sequences were processed as Little-Endian, Gentillon et al. BMC Res Notes (2016) 9:496 Page 6 of 18

Implicit and 16-bit uncompressed DICOM (RAW). All interest was a quest to assess quality in terms of actual, cap- BMP images were encoded with no down- sampling of tured details—free from illusionistic embellishments. In resolution. Both 3/1.5 T sequences were simultaneously other words, a DICOM file may appear as or show proper- exported as 16-bit DICOM, down-converted to 8-bit ties display of a 16-bit image on a computer; when in reality BMP (color space: RGB-24: 8 bits × 3 channels), and it only holds 12 bits of the real thing. Many of the measured batch-processed. Note that MicroDicom applied equal, 16-bit DICOM images contain much less than 65,536 tonal least-significant-bit (LSB) degradation to both 3/1.5 T patterns per pixel. It is due to the fact that not all allocated BMP images. The quality and the information loss in bits contain real captured details but bogus data. Reported both 3/1.5 T images were assessed with MaZda, a texture usable bit can be simply obtained with any software which analysis software which measures image statistics in vari- has a DICOM header viewer or editor. In this experiment, ous file formats [4, 5, 55]. Histogram is the only feature reported usable bits (in the DICOM headers) were first in MaZda version 5, to numerically measure available checked with MicroDicom. Then these metadata were dynamic range in 16-bit images. Histogram parameters double-checked and subsequently measured with MaZda were then used to calculate focus index values (skew- to determine actual stored bit in every image. Measured ness-to-mean kurtosis ratio) and dispersion index val- usable bit does not always correspond to reported usable ues (variance-to-mean ratio), which in turn serve as bit. Such discrepancies resulted from rounding errors in tools to evaluate the retention of stored, tonal details in MR units. Thus the choice to go with measured usable bit BMP ROIs—compared to the same ROIs in the original was factually validated (Fig. 8; Additional file 1: Dataset DICOM file (Figs. 4, 5, 6, 7, 8; Additional file 1: Dataset 1). Images with 4–5 mm thickness and highest measured 1). MaZda measured nearly 16 and 12 usable bits in 3 and usable bits were selected. For texture segmentation, coro- 1.5-T DICOM images, respectively (Fig. 8; Additional nal plane was ideal because more images could be obtained file 1: Dataset 1). No significant difference was observed from sequences with all four visible ROIs (i.e. thalamus, when testing was conducted with degraded BMP images ventricles, grey matter and white matter). Furthermore it (max. 8 bits in both) (Fig. 8; Additional file 1: Dataset 1). is worth mentioning that histogram and wavelets meas- Usable bit (stored bit) was used as a parametric factor urements depend on MRI signal types: e.g. “spin–lattice” rather than allocated or high bit, as the primary research relaxation time (T1 or T1 weighted image); “spin–spin” relaxation time (T2 or T2 weighted image); proton density (PD) (see Additional file 2: Dataset 2, Additional file 3: Dataset 3 for additional details).

Index of focus Focus index or index of focus (abbreviated: I-focus) is neither an integral parameter in MaZda nor in B11. It has been scarcely used before to measure perceived quality and realness of wood images [56], an embellish- ing technique for enhancing appearance of floor murals, backdrops and wall decorations used in low-cost construction. Focus index is calculated by dividing skewness by kurtosis. Skewness is a parameter which measures surface symmetry or lack of symmetry, imperfections, scanner misalignment. It indicates the even portion of a surface and the direction of distortion in the uneven por- tions (negatively and positively skewed). On the other hand, kurtosis is a measurement of the peakedness and flatness of fine details and edges where textural variations occur. When an image is in focus, edge elements or objects are sharp. Focus index combines these two parameters to thus assess quality by measuring focal or perspective distortion and sharpness. An image may be Fig. 7 a 1.5 T MR image of a fetal brain: 12-bit DICOM format, coronal sharp but distorted and vice versa. In magnetic resonance section. b 3 T MR image of a different subject: 16-bit DICOM, coronal imaging, focus index varies from one perspective plane to plane. c Same image shown in “a” after conversion to BMP. d Same image shown in “b” after conversion to BMP another (i.e. axial, coronal, sagittal). Ideal focus index is zero or close to zero [56, 57]. Sharpness quality for focus Gentillon et al. BMC Res Notes (2016) 9:496 Page 7 of 18

Fig. 8 Measured usable bits (pixel values). 3/1.5 T BMP 256; 1.5 T DICOM 4096; 3 T DICOM 65,536 ≈ ≈ ≈ index values outside the range of [−1, +1] can be eas- several other factors affect the visual interpretation of ily perceived by the naked, normal human eye. It is also the MR image which reaches the radiologist via medical- important to note that focus index alone does not deter- diagnostic LCD (e.g. acutance, physical and technical mine image quality. The common pitfall of the skewness unit settings, SNR, FOV, extended knowledge of elements and kurtosis is their sensitivity to artifacts like noise, affecting perceived appearance, realness, sharpness, etc) grain, post-editing sharpening and Gaussian blur [56, 57]. [59]. “Sharper” does not always signify “more details” and “more details” does not always signify “sharper.” Quality Index of dispersion depends on more factors than just “more details”. A MR Index of dispersion [also known as variance-to-mean image can have “higher resolution” and “more details” ratio (VMR)] is a measurement used to determine the and be not sharp or be degraded by artifacts and thus of clustering or dispersion of a luminance. In this research, poor quality. Unless its quality can be restored with com- VMR was utilized to quantify the volatility (difference) puter vision and/or editing tools, such an image would be of each individual ROI (ventricle, thalamus, grey matter, useless for live medical application. In clinical practice, white matter). The larger the difference between the coef- MR units have been used to detect anatomical, functional ficients of dispersion (dispersion index values) the greater and molecular anomalies. MR imaging modalities and the variability between the ROIs. Variance is the VMR the quality of their products are crucial because physi- component which measures contrast and boundaries cians not only rely on them to reach a firmly conclusive (edges) in a surface. Like kurtosis and skewness, variance diagnosis but also to make final interpretation to admin- is also affected by noise [58]. Variance is the difference in ister therapeutic care; which if wrong can lead to mal- luminance and/or color that makes an object or its rep- practice litigation. resentation in an image or a ROI to display distinguish- able. In other words, variance is a histogram parameter Fisher coefficient which can be used to determine contrast volatility in an One of the most promising application of texture analy- image. Finally, mean is a reflection of the average gray- sis in medicine is early detection of tumorous signs and scale luminance in MR images. It is directly proportional prophylaxis of cancer. A methodology of interest to this to the stored bits of an image. research was Fisher texture analysis. It has been used before in medical research to discriminate healthy from Resolution benign or malignant tissue. In this experiment, it was nec- 3 T scanner is by default constructed to capture more essary to develop a modified calculation of Fisher coeffi- details than 1.5 T. Its higher magnetic strength, when cient—coined as “feature-selection coefficient” (Eq. 1). It used properly, can also overcome problems with signal- was derived from an amalgamation of Fisher coefficient to-noise ratio (SNR). Therefore, 3 T units can condi- and statistical principles of analysis of variance (ANOVA). tionally produce image with more information and less 1 K K − 2 2 2 =1 =1 P Pjµ µj distortion. This superiority is technically due to denser D 1− K P k j k k F = = k=1 k (1) sampling of K-space for the same size of field of view V 2 K 2 k=1 Pk V (FOV), resulting in increased resolution. Nonetheless, k Gentillon et al. BMC Res Notes (2016) 9:496 Page 8 of 18

In Eq. 1, feature-selection coefficient is derived from images (Additional file 1: Dataset 1); but ROI discrimina- Fisher coefficient and ANOVA, where D is between-class tion was still better in 3 T. In both pre- and post- com- variance, V within-class variance, Pk probability of fea- pression, F was larger for 3 T. Unexpectedly, F was even ture k, Vk variance value of feature k in given class and better in the degraded images. μk mean value of feature k in given class (see MaZda user manual, http://www.eletel.p.lodz.pl/mazda/download/ Supplemental trial runs mazda_manual.pdf for further details). A preliminary trial was carried out as per recommenda- Resolution already indicated that 3 T captured more tions in MaZda and B11 user manuals/tutorial guides details than 1.5 T. However, it was still crucial to test (Fisher coefficient computation). The manufacturer’s -rec the differential capability of each modality in their own ommended method was tested with two different frames defined three-dimensional (3D) space (with disper- (images), extracted from a single 3 T sequence and two sion index, focus index, wavenhl) and the effects of different frames from a single 1.5 T sequence. Then two image compression on individual parameters. Features overlapping ROIs were selected and grouped—per ana- extracted from MaZda histogram were plotted on B11 tomical structures in each image (i.e. 2 ROIs of thalamus, XYZ space along with two controls [3T: Cmin (0, −1, 2 ROIs of ventricles, and so on). The outcome is delin- 0), Cmax (5500, 1, 120,000); 1.5 T: Cmin (0, −1, 0), Cmax eated later in this section herein. It is also important to (307, 1, 120,000)]. Note that only the x-axis is different. note that some aspects of this research were out of the This comparative method ensured that 1.5 and 3 T are researchers’ control—meaning that radio-technicians fairly measured within the resolution limit of the tested performed MRI examinations that attending physi- samples. The B11 program computed F from ANOVA cians and hospitals deemed necessary and safe for their derived parameters (F = D2/V2, where D = between- patients. Thus the images collected were limited to the class variance and V = within-class variance). Then it setting modes delineated in the MRI prescriptions and placed ventricles, thalamus, grey matter, and white mat- hospital-approved imaging protocols—which in turn is ter within a 3D space. By themselves, Cmin and Cmax contingent upon routine checkups, suspected patholo- yielded a maximum F of 10E + 6 and a misclassified gies and gestational age—as well as regional-clinical data error (MDE) of 0% for both 1.5 and 3 T. The effect practice. The latter varies from one location to another. of sample-mismatching on F and the software mechan- In the USA, for example, Food and Drug Administration ics were closely studied with control parameters and 3D (FDA) does not approve MRI with intravenous gadolin- graphs. It was observed that changing one Cmax to [0, −1, ium-based contrast agents for use during pregnancy— 60,000] while others remain constant caused a 25% MDE although nobody has yet discovered any hazards to fetus. and F value dropped to 18. A slice thickness of 3 mm is commonly used for routine fetal brain examinations, while 4-mm is used for other Results organs. In all collected samples, magnetic resonance Wavenhl, focus index (skewness-to-kurtosis ratio) and imaging was performed as an adjunct to ultrasound for dispersion index (variance-to-mean ratio) reveal better clarification of conditions mostly not related to brain quality for 3 T (Fig. 9). Though both 1.5-T and 3-T images anomalies. Throughout this research, the selection con- were 16-bit DICOM encoded, nearly 16 and 12 usable sisted of 1.5 and 3 T sequences which were within the bits were measured in 3-T and 1.5-T images, respectively. acceptable specifications: no apparent brain defects; Four bits in all 1.5 T images were padded. Such K-space slice thickness: 3–5 mm; repetition time: 800–2100; echo encoding methods appeared to reduce noise, by add- time: 1–95; imaging frequency: 60–130; pixel bandwidth: ing illusionistic details which are not really part of the 400–1200. Final sorting step necessitates pixel-by-pixel image. In contrast, all 3 T images were zero-bit padded. processing of 16-bit grayscale and then identification of This encoding technique provides space for storing more parameters with lowest standard deviation. What does details and increases the likelihood of noise but as well this sorting algorithm entails. It means that all MaZda as edges—which in turn are very crucial for differentia- parameters limited to 8-bit and 12-bit feature extraction of closely related anatomical structures. Both encod- tion were excluded. From the parameter list matching ing modes are possible with both units, but higher 3 T the criteria, three (skewness, kurtosis, and mean) were resolution is the main difference. Apart from surprisingly chosen for the B11 tests because the measurements were larger Fisher coefficient (Fig. 9), no significant paramet- fixed every time the process was repeated and had low- ric difference was observed (p > 0.05) when DICOM files est standard deviations between frames. Furthermore with 12 and 16 stored bits were degraded to 8-bit BMP skewness, kurtosis, and mean parameters had measure- (Additional file 2: Dataset 2). MaZda measured an equal ments that were very close to the average values. These detrimental loss of quality in both 1.5-T and 3-T BMP extracted features were then analyzed in 3D space with Gentillon et al. BMC Res Notes (2016) 9:496 Page 9 of 18

Fig. 9 Graph showing difference between ROIs. Raw-data analysis was performed to compute F with 1-nearest-neighbor (NN) classification and no feature standardization. Same samples and ROIs were used in both pre- and post- image compression. a 1.5 T uncompressed DICOM; Fisher coefficient computation with controls: Cmin (0, 1, 0); Cmax (307, 1, 120,000); F 426.0; MDE 0%. b Same 1.5 T samples: zoomed in, mostly = − = = = on the y-axis. c same 1.5 T after compression to 8 bits; Fisher coefficient computation with controls: Cmin (0, 1, 0) Cmax (30, 1, 120,000); = − = F 776.0; MDE 5.56%. d 3 T uncompressed DICOM: F 1787.0; MDE 0%. e Same 1.5 T samples: zoomed in, mostly on the y-axis. f Same 3 T = = = = after compression to 8 bits. 3 T: F 2344.3; MDE 0% = =

B11. For images extracted from the same 1.5 and 3 T Different sequences from different patients were scored sequences, RAW analysis yielded a maximum Fisher (Fig. 11). Fisher coefficient was 897.4 for 3 T and 144.1 coefficient of E + 5 and perfect 0% misclassification for 1.5 T. It is important to note that proton density (PD) error for both imaging modalities. B11 plugin was used sequence is very different from T2. The trial was re-run to generate and examine the result on a graph (Fig. 10). with two 1.5 T T2 HASTE (half-Fourier acquisition sin- Though the graph appears as texture analysis of a single gle-shot turbo spin-echo) sequences from two different image, each number is actually from two different frames patients; and F was 448.6. The software maker did not overlapping in three-dimensional space. The software have an immediate solution for the B11 limitation and guidelines as well as its engineers were consulted regard- thus recommended to use the software when compara- ing B11 limitation. It turned out that B11’s Fisher Coef- tively testing many images from different sequences. ficient (F) is internally computed from modified ANOVA In technically controlled settings, the software maker statistics—by customizedly dividing a nominator (dif- suggestions are feasible. However, in medical practice, it ference between regions) over a denominator (differ- is not a viable process. It would be detrimental to get so ence within regions). We did not investigate further and many MR studies from a single patient for ROI analysis. neither modify the computation mechanics of Fisher Coefficient, as the software license did not authorize us Grayscale quantification with histogram‑based parameters to do so. Larger F was construed to mean that there is There are misconceptions about the usefulness of gray- more difference between ROIs. For the purpose of this scale quantification and its applications in medicine. research, straight-forward comparative test with Fisher While this area of diagnostic radiology is still emerging coefficient method was inconclusive, as F was the same and under-documented—grayscale texture analysis with for comparison between 3 and 1.5 T ROIs (see Additional histogram-based parameters is already proven to be very file 3: Dataset 3—appendix 1, appendix 2, appendix 3 for useful for differentiation of benign from malignant thy- further details). roid cancer, for example. Such parameters can be used The software maker recommended to use different to detect micro-calcifications inside nodules—which is frames from different MR sequences and/or studies. an early sign of thyroid malignancies. In this research, Gentillon et al. BMC Res Notes (2016) 9:496 Page 10 of 18

Fig. 10 a Difference between ROIs for 3 T with two images from same T2 sequence (same patient). 1 ventricle; 3 thalamus; 2 grey matter; 4 white matter. b Difference between ROIs for 1.5 T with two images from same T2 sequence (same patient). 1 ventricle; 3 thalamus; 2 grey matter; 4 white matter

Fig. 11 a Difference between ROIs for 3 T with two images from different T2 TSE [Turbo spin echo sequences (different patients)]. 1 ventricle; 3 thalamus; 2 grey matter; 4 white matter. b Difference between ROIs for 1.5 T with one image from PD and one from T2 HASTE sequence (different patient). 1 ventricle; 3 thalamus; 2 grey matter; 4 white matter

histogram parameters was used to calculate focus index grain (simulated sharpness), peakedness and flatness of ratios (skewness-to-mean kurtosis ratio) and dispersion fine details and edges (due to image resolution), chro- index values (variance-to-mean ration). What is the logi- matic aberration blur (optical distortion) and other arti- cal meaning of using such parameters in magnetic reso- facts. Focus index would probably sounds novel to most nance imaging? Histogram-based parameters don’t just or perhaps all readers of this article. Yet, it is sparsely used generically measure intensity of grayscale and signal-to- in reality-rendering imaginary to measure realness (natu- noise ratio (SNR). Such parameters can also be used for ralness) of surface perception—and thus the phenome- morphometric measurements, texture segmentation and non that human eye does not always sees what it thinks it discrimination. The functionality of focus index can be does. Applications of focus index include photo-realistic simply understood from its numerator and denominator. rending, optical/visual illusion, methatetic continuum In computational- visual cognition and imaging statistics, (change in stimulus quality) and indiscernible visual per- skewness is a numerical measurement of symmetry and ception between natural materials and replicas (e.g. out- dissymmetry in an image. Kurtosis, on the other hand, is door backdrops, floors and murals) used in filmmaking a histogram parameter that is sensitive to Gaussian blur and architectural construction. A pitfall with focus index and out-of-focus blurs (indirect sharpness), noise and is its susceptibility to digital and mechanical artifacts Gentillon et al. BMC Res Notes (2016) 9:496 Page 11 of 18

recorded by imaging equipment but not really part of the Hence the larger variance observed in 3 T parametric image. How was the problem solved? Perceived quality of measurements are likely due to sharply captured ana- photo-realistic prints was previously documented to be tomical details with fine edges—rather than digitally and optimally construed as real—when focus index is zero or mechanically generated noisy artifacts. nearly zero but within −1 and +1 interval. Furthermore focus index and dispersion index were concomitantly Computation of difference between ROIs used to rule out noise (artifacts) from true signal (actual It would be subjective to just judge the quality by simply anatomy). The definition of dispersion index can be tech- comparing 1.5 and 3 T images on histogram graphs. As nically interpreted in the difference between its dividend a remedy, other comparative tests were developed and and divisor as well as the resulting quotient. Variance is a carried out. In Fischer coefficient, the numerator (differ- numerical measurement of contrast and captured details ence between the ROIs) was the quantification wanted. in an image; and thus it is affected by resolution. Mean Hence analysis of variance was subsequently performed is the average luminance derived from stored bit value to quantify the difference between 3 and 1.5 T ROIs. in an image (ROI in this case). Mean may be misunder- The parameters were then imported into STATISTICA stood as a mere measurement of pixel values, but herein version 10 for further statistical processing and analy- it also served as a reference value for the variance. Stored sis. All the results were exported to excel, and a dataset bit is not to be confused with allocated bit and high bit. was generated (Additional file 4: Dataset 4). 3 T images The stored bit value reported in the DICOM tag did not (~E + 15) had much higher ROI variations than 1.5 T always correspond to that measured with MaZda. A large images (~E + 10). In terms of parametric quantifica- gap between the variance and the mean could be an indi- tion, the increased variability between ROIs revealed that cation of fine details or edges (e.g. anatomical bounda- 3 T is a better discriminative tools than 1.5 T (Fig. 12). ries or pattern changes within anatomical structures); Focus index values and dispersion index values were used or it could also be due to randomly occurring noise and/ respectively to measure sharpness distribution and sta- or unwanted artifacts. Both 1.5 and 3 T images were tistical dispersion (Figs. 13, 14, 15, 16). The focus index encoded in 16-bit DICOM container. Why don’t they graph shows that 3 T images have better sharpness and have the same quality? Not all bit slots necessarily con- better focus in both groups. As per dispersion index tain true details (actual anatomy). In our test, converting graph, 3 T is also better in terms of spatial resolution. In (uncompressing) 16-bit Lossless JPEG DICOM to 16-bit all measurements, the results were statistically significant RAW DICOM did not improve quantification of bit slots with p < 0.01 (see Table 2). which actually contain anatomical details—though the RAW file was larger. If stored bit was initially 12, it was Discussion still 12 after uncompression. MR scanner settings are In this experiment, the findings revealed that 8-bit also known to digitally and mechanically affect quality. BMP actually yielded higher F than the original 16-bit Therefore, 3 and 1.5 T images were closely matched by DICOM. Larger F coefficient does not necessarily mean matrix size, field-of-view (FOV), slice thickness, phase better quality. Nonlinear discriminant analysis (NDA), and frequency encodings. Besides magnetic strength and linear discriminant analysis (LDA), and principal com- speed, the quality of captured details also depends on ponent analysis (PCA) can generate low or large F the voxel size and thus the resolution of 1.5 and 3 T (in when B11 is overtrained [60]. Thus RAW data analysis the collected samples, 256 × 256 and 446 × 446 respec- was performed, as it is more constant than NDA, LDA, tively). Separating noise and discretization artifacts from and PCA transformations. With featured standardiza- true signals is a time-consuming editing process. A more tion turned-off, RAW data analysis calculated same F efficient solution was to get rid out of heavily noisy sam- values each time measurements are repeated. In this ples. So doing does not necessarily imply that we think case, larger F from 8-bit BMP is unlikely to be due to that noise is always bad. Despite its degradative nature, overtraining problem in the network. It is rather due noise can also improve visual appearance of an image. to squeezing tricks employed in lossy compression. Dither, for instance, is in-machine (real-time) or post- While ROIs appear more uniform, some fine details editing added noise to filter some unwanted artifacts get discarded. In MaZda texture analysis software, his- (e.g. off-resonance bandings, posterization)—in order to togram can indeed measure regional tonality directly improve acutance (sharpness) and therefore visual per- from image statistics rather than from secondary vis- ception. By narrowing the reference range of the focus ual means such as graphical appearance on a monitor index quotient to [−1, +1], a large number of noisy sam- screen; which is susceptible to eye limitations (e.g. acu- ples were excluded, as a result of displaying character- tance, screen size, etc). Thus 16-bit image can be accu- istics of random quantum mottle (grainy appearance). rately mapped three-dimensionally from intensity value Gentillon et al. BMC Res Notes (2016) 9:496 Page 12 of 18

Fig. 12 ROI variability graph showing difference between 1.5 and 3 T

of 0 to a maximum intensity of 65,535; and variance pro- matter). Concomitant texture analysis with wavelets and vides information about details, sharpness and edges in histogram allows accurate classification of closely related the image. 3 T is technically able to capture more details anatomical structures in the brain during development, due to higher resolution. Nevertheless, higher resolu- and thus an important set of tools for early detection tion does not always guarantee better quality and neither of tissue changes in the brain. For example, grey matter does it always mean more details. Without proper unit is vastly present in the brain. It can be difficult for the settings to booster available dynamic range, resolution unaided human eye to trace the boundaries of hypotha- is just a mere increment in the cost of storage media. lamic nuclei on a prenatal magnetic resonance image. It Available dynamic range of an image is the informa- is due to the fact that grey matter is also present deep tion that is not seen until it is actually used. The origi- inside the cerebrum and inside the thalamus, hypothala- nal encoding bit space is crucial for storage of captured mus, and the basal ganglia. Even with the fetal brain fully details. In 8-bit space, for instance, it is mathematically formed, it is still elusive for the human eye to deline- impossible to recover fine highlight details such as blue ate and differentiate these anatomical structures during sky in an indoor image of a blown-out window on a pregnancy. In the overexposure example, the human bright sunny day. Lowering exposure shows grey sky. In eye would see the exact same blown-out region with no 12- or higher-bit space, the sky should be still blue. This obvious quality discrimination in either 8-bit or 16-bit is where available dynamic range comes in play. Wave- space, while histogram and wavelets would pick up the lets, on the other hand, were not used for measuring hidden details and reveal a difference in parametric val- quality but for double-checking and matching regions of ues. This is where texture analysis software fits in bio- interest (i.e. ventricles, thalamus, grey matter, and white medical imaging. Gentillon et al. BMC Res Notes (2016) 9:496 Page 13 of 18

Fig. 13 Sharpness distribution: Group 1 acceptable range is [ 1, 1]. Zero is absolute sharpness; in other words, better focus. 3 T has more points + − closer to zero

Fig. 14 Sharpness distribution: Group 2 acceptable range is [ 1, 1]. Zero is absolute sharpness; in other words, better focus. 3 T has more points + − closer to zero

Constraints, limitations, and assumptions BMP, and it can measure the actual stored bits within MaZda Texture Analysis version 5 was used through- a DICOM container of 16 allocated bits. This feature is out this experiment (Additional file 5). It is a quantita- crucial in terms of comparing 1.5–3 T for discriminative tive analysis tool which can extract various parameters power of closely related anatomical structures in fetal from 16-bit DICOM without need of conversion to 8-bit brain imaging. Unfortunately, many MaZda parameters Gentillon et al. BMC Res Notes (2016) 9:496 Page 14 of 18

Fig. 15 Statistical dispersion: logarithmic graph (Group 1) showing difference between 1.5 and 3 T. 3 T ROIs are more spread out; thus better for dis‑ tinguishing close anatomical structures in fetal MRI. Regions of interest—red: ventricle—yellow: white matter—blue: thalamus—green: grey matter

Fig. 16 Statistical dispersion: logarithmic graph (Group 1) showing difference between 1.5 and 3 T. 3 T ROIs are more spread out; thus better for dis‑ tinguishing close anatomical structures in fetal MRI. Regions of interest—red: ventricle—yellow: white matter—blue: thalamus—green: grey matter

were excluded from this experiment because of either Recommendations 8-bit or 12-bit limitations. An updated, customized ver- As of today, there are still numerous talks about the sion of this software package would require several man- usefulness of 3 T MRI in medical practice and the pos- hours and therefore was not feasible in due time. sibility of deleterious effects during early pregnancy. The Gentillon et al. BMC Res Notes (2016) 9:496 Page 15 of 18

Table 2 The best nine values were selected to determine p values

IFOCUS IDISPERSION 1.5 T G1 3 T G1 1.5 T G2 3 T G2 1.5 T G1 3 T G1 1.5 T G2 3 T G2

0.174816 0.001664 0.261094 0.013746 245.8439 3845.506 272.6785 3919.487 0.155916 0.001614 0.076755 0.001496 241.6155 3725.265 208.9656 3684.96 0.07019 0.000614 0.002159 0.000435 222.8727 3584.081 207.59 3605.269 0.084207 0.000317 0.010039 5.16E 07 183.2519 3518.256 191.5476 3396.184 − 0.12892 0.000824 0.018455 3.76E 06 152.8732 3234.123 62.18858 1969.419 − 0.14576 0.001041 0.053909 0.00046 63.68822 1949.152 59.11231 1557.758 0.222427 0.001126 0.095316 0.00049 62.51141 1861.574 56.22063 1547.73 0.275593 0.002816 0.100121 0.004151 52.44858 1775.027 50.28433 1542.073 0.315591 0.003521 0.239347 0.01374 48.76635 1763.935 28.30434 614.1261 m 0.175 m 0.0015 m 0.0952 m 0.0038 m 142 m 2806 m 126 m 2426 ======sd 8E 02 sd 1E 03 sd 9E 02 sd 5E 03 sd 85 sd 935 sd 92 sd 614 = − = − = − = − = = = = t(16) 6.3 t(16) 2.9 t(16) 8.5 t(16) 5.6 p 0.0001= p 0.011= p 0.0001= − p 0.0001= − = = = =

scientific views about the differential quality between was not necessarily an advantage. Yet the World Health 1.5 and 3 T are mixed and perplexed due to subjectivity Organization (WHO) has noticed a fall in misdiagnosis and unreliable interpretative techniques, which largely and treatment cost and a rise in detection and diagnosis depend on the human eye perception. Observational of the pandemic disease of tuberculosis (TB)—with intro- studies have reported that the greater magnetic field of duction of computer-aided interpretation software in 3 T produces sharper images based on visual appearance. regions with prevalent TB outbreak. Again, this research Such findings are subjected to controversion because is not about whether human vision is better than artifi- different pairs of eyes don’t always see the same thing. cial vision or vice versa. We disclaim any so-construed The normal human eye was reported to have a differ- allegations. The goal is that texture analysis may assist ential power of a quasi 8-bit scanner. Very few people obstetricians and radiologists in making more accurate perceive details beyond that limitation. From a financial and objective medical diagnosis of prenatal pathologies. point of view, billions of dollars could be saved if 1.5 T is Reaching this outcome is a matter of testing and finding used instead of 3 T MRI. Nevertheless, financial stand- suitable artificial systems and methods of texture analysis point and hardship may indeed prevent a clinician from for the appropriate set of images. Last but not least, arti- ordering 3 T MRI, instead of using relevance to benefi- ficial magnetic fields of 1.5 T (30,000 times greater than ciaries as a guide to make such a decision. Some still raise earth’s magnetic field) and 3 T (60,000 times greater than questions over the safety of 3 T magnetic field in medi- earth’s magnetic field) [61], scanners have raised theoreti- cal practice. Nevertheless, twelve years ago, studies have cal concerns in the medical community. Unknown risks already shown that 8T MRI causes no obvious damage such as teratogenic and biological effects—if they really to the human body. Like other imaging modalities, the exist—could be reduced by using 1.5 T during the first main issue in the medical community that is affecting 28 weeks—as its quality is sufficient [62–67]. the interpretation of 3 T MR images is subjectivity and observer dependence. It is a challenging task for different Conclusions human eyes to correctly identify square-pixel variations Unquestionably, 3-T exhibits better quality than 1.5-T between the gray-scale image presented in Fig. 1. A key fetal magnetic resonance imaging. The results were sig- feature of texture analysis with computer vision is that it nificant (p < 0.05). Nevertheless, 1.5-T is of sufficient can quantify texture in pixelated 1.5 and 3 T MR images quality for routine fetal examination during early preg- into numerical values, which in turn can be used to assess nancy and therefore can lower the risks of unknown tera- image quality—therefore eliminating subjectivity and togenic effects. Though 3 T fetal MRI exhibits superior reproducibility issues in diagnostic imaging. Some of our image quality, its usefulness, health claims, side effects, guest reviewers had argued that texture quantification benefits, and safety demand further investigation. Gentillon et al. BMC Res Notes (2016) 9:496 Page 16 of 18

Additional files of the MaZda/B11 software and neither financed nor encouraged by COST. Hence it is not intended to be a promotion for COST.

Additional file 1: Dataset 1. Figure 8 data: stored bits in 1.5/3 T DICOM/ Availability of data and materials BMP, measured with MaZda version 5. Links to download raw data are included in this manuscript herein. To get access to medical records, individual solicitor must send formal request to the Additional file 2: Dataset 2. Effects of pull-down conversion: 12/16-bits research committee of the Medical University of Lodz. grayscale DICOM to 8-bit BMP (RGB-24). Recommended resources: MaZda user manual (http://www.eletel.p.lodz. Additional file 3: Dataset 3. Raw texture analysis/Fisher coefficient: ➤ pl/mazda/download/mazda_manual.pdf); MaZda tutorials (http://www. appendix 1, ➤ appendix 2, ➤ appendix 3. eletel.p.lodz.pl/programy/mazda/index.php?action tutorial); MaZda Package v4.6 (http://www.eletel.p.lodz.pl/programy/mazda/=); textbook: “Texture analy‑ Additional file 4: Dataset 4. Data for supplemental trial runs: 36 1.5 T; sis for magnetic resonance imaging.” 234 pages, ISBN: 80-903660-0-7 Publisher: 36 3 T; group 1: 20–28 weeks; group 2: 29–40 weeks. × × Med4publishing s.r.o., 2006. Additional file 5. MaZda Package v5 RC HG (release candidate: Hugues Gentillon): this pre-release (beta) version contains: 1) MaZda version 5 for Consent to publish achromatic and chromatic image analysis, 3D editing, feature extraction Consent for publication was obtained from patients. Furthermore all data from & selection, geometrical features for ROIs; 2) B11 version 3.3 for feature human participants were anonymized as per informed consent agreement. clustering, image segmentation, unsupervised & supervised classification. Bioethics Committee Approval Number: RNN/213/13/KE, July 16, 2013.

Ethics approval and consent to participate Abbreviations The authors declare that written informed consent was obtained from all 2D: two-dimensional; BMP: bitmap; CNS: central nervous system; ROI: region patients who were enrolled in the study. Ethics approval was granted by the of interest; CAD: computer-aided diagnosis; COST: cooperation in science research committee at the Medical University of Lodz. Agreement number: and technology; DICOM: digital imaging and communications in medicine; 3/2011—concluded on December 6, 2011, between the experimenter E : positive exponent; E-: negative exponent; JPEP: Joint Photographic (Hugues Gentillon), research supervisor (Ludomir Stefańczyk) and rector of Experts+ Group; LCD: liquid crystal display; HG: Hugues Gentillon; LS: Ludomir Medical University of Lodz (Radzisław Kordek), Faculty of Biomedical Sci‑ Stefańczyk; MS: Michał Strzelecki; MRL: Maria Respondek-Liberska; MR: mag‑ ence—Post-graduate research in Diagnostic Imaging and Radiotherapy. netic resonance; MRI: magnetic resonance imaging; PD: proton density; RAW: reas as written; RGB: red, blue, green; RGB-24: color space with 8 bits per chan‑ Funding nel (8 8 8); ROI: region of interest; US: ultrasound; USG: ultrasonography; T: Self-funded; material supports from the Medical University of Lodz/Polish × × Research Committee and affiliated institutions and hospitals; grants and finan‑ tesla; TB: tuberculosis; VROI: variations between ROIs; Wk: week. cial aid from Swedish Ministry of Education and Research/Centrala Studiestöd‑ Authors’ contributions snämnden and from U.S. Department of Education. HG conceived, designed, and carried out the experiment, interpreted the data and drafted the manuscript. Both LS and MRL provided human samples and Received: 29 February 2016 Accepted: 15 November 2016 participated in conceiving/designing the clinical aspects of the research and revising its manuscript. MS provided texture analysis software/plugins and par‑ ticipated in conceiving/designing the technical aspects of the experiment and in revising its manuscript. All authors read and approved the final manuscript.

Author details References 1 Department of Radiology and Diagnostic Imaging, Barlicki University 1. Hellman LM, Duffus GM, Donald I, Sunden B. Safety of diagnostic ultra‑ Hospital, Medical University of Lodz, Lodz, Poland. 2 Institute of Electronics, The sound in obstetrics. Obstet Gynecol Surv. 1971;26(1):13–4. Faculty of Electrical, Electronic, Computer and Control Engineering, Technical 2. Richards DS. Prenatal ultrasound to detect fetal anomalies. NeoReviews. University of Lodz, Lodz, Poland. 3 Diagnosis and Prevention of Congenital 2012;13(1):e9–19. Malformations, Instytut Centrum Zdrowia Matki Polki, Lodz, Poland. 3. Nishi T, Iwasaki M, Yamoto M, Nakano R. Prenatal diagnosis of autosomal recessive polycystic kidney disease by ultrasonography and magnetic Acknowledgements resonance imaging. Acta Obstet Gynecol Scand. 1991;70(7–8):615–7. The authors gratefully acknowledge prof. Rafał Pawliczak and MUL staff (Divi‑ 4. Kawabata I, Takahashi Y, Iwagaki S, Tamaya T. MRI during pregnancy. J sion of Doctoral Studies for International Students, Biomedical Sciences, Medi‑ Perinat Med. 2003;31(6):449–58. cal University of Lodz) for research coordination, logistics, and administration; 5. Bulas D, Egloff A. Benefits and risks of MRI in pregnancy. Semin Perinatol. prof. Paweł Liberski and MUL (Division of Doctoral Studies for International 2013;37(5):301–4. Students, Biomedical Sciences, Medical University of Lodz) neuropathology 6. McKenna DA, Meehan CP, Alhajeri AN, Regan MC, O’Keeffe DP. The use department for counsels with funds to cover MRI expenses; prof. Ludomir of MRI to demonstrate small bowel obstruction during pregnancy. Br J Stefańczyk and Barlicki hospital staff (Department of Radiology and Diagnostic Radiol. 2014;80(949):e11–4. Imaging, Barlicki University Hospital, Medical University of Lodz) for sample 7. Curtis M, Hopkins MP, Zarlingo T, Martino C, Graceansky-Lengyl M, Jeni‑ supply and clinical feedback; prof. Maria Respondek-Liberska and Matki Polki son EL. Magnetic resonance imaging to avoid laparotomy in pregnancy. Hospital (Diagnosis and Prevention of Congenital Malformations, Instytut Obstet Gynecol. 1993;82:833–6. Centrum Zdrowia Matki Polki, Lodz, Poland) for sample supply and clinical 8. Quinn TM, Hubbard AM, Adzick NS. Prenatal magnetic resonance imag‑ feedback—as well as prof. Tadeusz Biegański, ICZMP director; prof. Michał Str‑ ing enhances fetal diagnosis. J Pediatr Surg. 1998;33(4):553–8. zelecki, Sławomir Skonieczka and TUL staff (Institute of Electronics, The Faculty 9. De Wilde JP, Rivers AW, Price DL. A review of the current use of magnetic of Electrical, Electronic, Computer and Control Engineering, Technical Univer‑ resonance imaging in pregnancy and safety implications for the fetus. sity of Lodz, Poland) for providing MaZda software and technical feedback. Prog Biophys Mol Biol. 2005;87(2):335–53. 10. Ertl-Wagner B, Lienemann A, Strauss A, Reiser MF. Fetal magnetic reso‑ Competing interests nance imaging: indications, technique, anatomical considerations and a The authors declare that there might be unforeseen academic/institutional review of fetal abnormalities. Eur Radiol. 2002;12(8):1931–40. competing interests—which the authors themselves are not aware of. Never‑ 11. McCarthy SM, Filly RA, Stark DD, Hricak H, Brant-Zawadzki MN, Callen theless, this research and its results were carried out independently and thus PW, Higgins CB. Obstetrical magnetic resonance imaging: fetal anatomy. not endorsed by and in no way influenced by the inventors/patent owner Radiology. 1985;154(2):427–32. Gentillon et al. BMC Res Notes (2016) 9:496 Page 17 of 18

12. Weinreb JC, Lowe TW, Santos-Ramos R, Cunningham FG, Parkey 33. Lawn SD, Mwaba P, Bates M, Piatek A, Alexander H, Marais BJ, Abuba‑ R. Magnetic resonance imaging in obstetric diagnosis. Radiology. kar I. Advances in tuberculosis diagnostics: the Xpert MTB/RIF assay 1985;154(1):157–61. and future prospects for a point-of-care test. Lancet Infect Dis. 13. Huisman TA, Martin E, Kubik-Huch R, Marincek B. Fetal magnetic reso‑ 2013;13(4):349–61. nance imaging of the brain: technical considerations and normal brain 34. Blehm C, Vishnu S, Khattak A, Mitra S, Yee RW. Computer vision syndrome: development. Eur Radiol. 2002;12(8):1941–51. a review. Surv Ophthalmol. 2005;50(3):253–62. 14. Hartwig V, Giovannetti G, Vanello N, Lombardi M, Landini L, Simi S. Bio‑ 35. Livingstone M, Hubel DH. Vision and art: the biology of seeing (vol 2), logical effects and safety in magnetic resonance imaging: a review. Int J chapter 12. New York: Harry N. Abrams; 2002. Environ Res Public Health. 2009;6(6):1778–98. 36. Watson AB, Null CH (1997) Digital images and human vision. Nasa Res 15. Kok RD, de Vries MM, Heerschap A, van den Berg PP. Absence of harmful Center. 1997;20020041928:EI’97. effects of magnetic resonance exposure at 1.5 T in utero during the 37. Graps A. An introduction to wavelets. IEEE Comput Sci Eng. third trimester of pregnancy: a follow-up study. Magn Reson Imaging. 1995;2(2):50–61. 2004;22(6):851–4. 38. Akay M. Wavelet applications in medicine. IEEE Spectr. 1997;34(5):50–6. 16. Baker LC, Wheeler SK. Managed care and technology diffusion: the case 39. Bharati MH, Liu JJ, MacGregor JF. Image texture analysis: methods and of MRI. Health Aff. 1998;17(5):195–207. comparisons. Chemometr Intell Lab Syst. 2004;72(1):57–71. 17. Abe H, Schmidt RA, Shah RN, Shimauchi A, Kulkarni K, Sennett CA, 40. Castellano G, Bonilha L, Li LM, Cendes F. Texture analysis of medical Newstead GM. MR-directed (“Second-Look”) ultrasound examination for images. Clin Radiol. 2004;59(12):1061–9. breast lesions detected initially on MRI: MR and sonographic findings. Am 41. Hendrick RE. Artifacts and errors in breast magnetic resonance imag‑ J Roentgenol. 2010;194(2):370–7. ing. Breast MRI: fundamentals and technical aspects, ch 11. 2008. p. 18. Carmichael IW, MacLeod AM, Travlos J. MRI can prevent unnecessary 187–207. arthroscopy. J Bone Joint Surg Br. 1997;79(4):624–5. 42. Herbst M, Maclaren J, Lovell-Smith C, Sostheim R, Egger K, Harloff A, et al. 19. Lubovsky O, Liebergall M, Mattan Y, Weil Y, Mosheiff R. Early diagnosis of Reproduction of motion artifacts for performance analysis of prospective occult hip fractures: MRI versus CT scan. Injury. 2005;36(6):788–92. motion correction in MRI. Magn Reson Med. 2014;71(1):182–90. 20. Naredo E, Collado P, Cruz A, Palop MJ, Cabero F, Richi P, et al. Longitudinal 43. Somasundaram K, Kalavathi P. Analysis of imaging artifacts in MR brain power Doppler ultrasonographic assessment of joint inflammatory activ‑ images. Oriental J Comput Sci Technol. 2012;5(1):135–41. ity in early rheumatoid arthritis: predictive value in disease activity and 44. Zhuo J, Gullapalli RP. MR artifacts, safety, and quality control. Radiograph‑ radiologic progression. Arthritis Care Res. 2007;57(1):116–24. ics. 2006;26(1):275–97. 21. Battista R, Banta HD, Jonsson E, Hodge M, Gelband H. Lessons from the 45. Kohlenberg L. Everybody knows this is no wear. Am Journalism Rev. eight countries. In: Health care technology and its assessment in eight 1994;16(4):38–42. countries, OTA-BP-H-140, chap 7. US Congress, Office of Technology 46. Cromey DW. Avoiding twisted pixels: ethical guidelines for the appropri‑ Assessment. Washington, DC: US Government Printing Office, DIANE ate use and manipulation of scientific digital images. Sci Eng Ethics. Publishing. 1995; p. 209–58. 2010;16(4):639–67. 22. Girouard N, Imai Y. The health care system in Poland. OECD, Economics 47. Shotton J, Johnson M, Cipolla R. Semantic texton forests for image Department Working Papers. 2000. p. 257. doi: 10.1787/265022143317. categorization and segmentation. In: IEEE conference on computer vision 23. Lewis MA. Who is paying for health care in Eastern Europe and Central and pattern recognition, CVPR 2008; 2008. p. 1–8. Asia?, vol. 20940. Washington: World Bank Publications; 2000. p. 1–31. 48. Pal NR, Pal SK. A review on image segmentation techniques. Pattern 24. Materka A, Strzelecki M. Texture analysis methods—a review, Techni‑ Recogn. 1993;26(9):1277–94. cal University of Lodz, Institute of Electronics. COST, Brussels. 1998;B11 49. Zakrzewski M. Polish doctors implanted the world’s smallest artificial car‑ report:1998. diac pacemaker. University Hospital of Poznan, EPA. 2016; Press Release. p. 25. Szczypinski P, Strzelecki M, Materka A, Klepaczko A. MaZda - A software 52543790. package for image texture analysis. Comput Methods Programs Biomed. 50. Majda T. Polish cardiology started in Lodz 60 years ago—1950–2010. 2009;94(1):66–76. Cardiol J. 2011;18(1):110–3. 26. Strzelecki M, Szczypinski P, Materka A, Klepaczko A. A software tool for 51. Hrynaszkiewicz I, Norton ML, Vickers AJ, Altman DG. Preparing raw clini‑ automatic classification and segmentation of 2D/3D medical images. cal data for publication: guidance for journal editors, authors, and peer Nucl Instrum Methods Phys Res Sect A. 2013;702:137–40. reviewers. BMJ. 2010;340(8):c181. 27. Muyoyeta M, Maduskar P, Moyo M, Kasese N, Milimo D, Spooner R, 52. Rose A. Are face-detection cameras racist? Time magazine online. (http:// et al. The sensitivity and specificity of using a computer aided diagno‑ www.time.com/time/business/article/0,8599,1954643-1,00.html). 22 Jan sis program for automatically scoring chest X-rays of presumptive TB 2010;1954643-1. patients compared with Xpert MTB/RIF in Lusaka, Zambia. PLoS ONE. 53. Whitehill J, Bartlett MS, Movellan JR. Automatic facial expression recogni‑ 2014;9(4):e93757. tion. Social Emot Nat Artifact. 2013;ch6:88–9. 28. Maduskar P, Muyoyeta M, Ayles H, Hogeweg L, Peters-Bax L, van Ginneken 54. Frayer M. Giving the toaster eyes: applied anthropomorphism and its B. Detection of tuberculosis using digital chest radiography: automated influences on user-object relations with everyday objects. 2010; KOVM02 reading vs. interpretation by clinical officers. Int J Tuberc Lung Dis. 20101:id1604671. 2013;17(12):1613–20. 55. McKenna DA, Meehan CP, Alhajeri AN, Regan MC, O’Keeffe DP. The use 29. Breuninger M, van Ginneken B, Philipsen RH, Mhimbira F, Hella JJ, Lwilla of MRI to demonstrate small bowel obstruction during pregnancy. Br J F, et al. Diagnostic accuracy of computer-aided detection of pulmonary Radiol. 2007;80(949):e11–4. tuberculosis in chest radiographs: a validation study from sub-Saharan 56. Katsura S, Mizokami Y, Yaguchi H. Perceived quality of wood images Africa. PLoS ONE. 2014;9(9):e106381. influenced by the ratio of skewness to kurtosis of image histogram. The 30. Hillemann D, Rüsch-Gerdes S, Boehme C, Richter E. Rapid molecular 1st Asia color association conference, 2013, Proc. p. 86–9. detection of extrapulmonary tuberculosis by the automated GeneXpert 57. Katsura S, Mizokami Y, Yaguchi H. Perceived quality of wood MTB/RIF system. J Clin Microbiol. 2011;49(4):1202–5. images influenced by the skewness of image histogram. Opt Rev. 31. Steiner A, Mangu C, van den Hombergh J, van Deutekom H, van Gin‑ 2015;22(4):565–76. neken B, Clowes P, Hoelscher M. Screening for pulmonary tuberculosis 58. Law WK, Chung AC. Minimal weighted local variance as edge detector for in a Tanzanian prison and computer-aided interpretation of chest X-rays. active contour models. In: Computer vision–ACCV, 2006; Berlin: Springer; Public Health Action. 2015;5(4):249. 2006. p. 622–632. 32. Philipsen RHHM, Sánchez CI, Maduskar P, Melendez J, Peters-Bax L, 59. Schmitz BL, Aschoff AJ, Hoffmann MH, Grön G. Advantages and Peter JG, van Ginneken B. Automated chest-radiography as a triage for pitfalls in 3 T MR brain imaging: a pictorial review. Am J Neuroradiol. Xpert testing in resource-constrained settings: a prospective study of 2005;26(9):2229–37. diagnostic accuracy and costs. NPG Sci Rep. 2015;5:12215. doi:10.1038/ 60. Materka A. Texture analysis methodologies for magnetic resonance imag‑ srep12215. ing. Dialogues Clin Neurosci. 2004;6(2):243–50 Gentillon et al. BMC Res Notes (2016) 9:496 Page 18 of 18

61. Jerrolds J, Keene S. MRI safety at 3 versus 1.5 T. Internet J World Health 65. Whitby EH, Paley MNJ, Sprigg A, Rutter S, Davies NP, Wilkinson ID, Griffiths Soc Politics. 2009;6:1. PD. Comparison of ultrasound and magnetic resonance imaging in 100 62. Cannie M, Jani J, Dymarkowski S, Deprest J. Fetal magnetic reso‑ singleton pregnancies with suspected brain abnormalities. BJOG Int J nance imaging: luxury or necessity? Ultrasound Obstet Gynecol. Obs Gynaecol. 2004;111(8):784–92. 2006;27(5):471–6. 66. Altman RP, Randolph JG, Lilly JR. Sacrococcygeal teratoma: American 63. Levine D, Barnes PD, Robertson RR, Wong G, Mehta TS. Fast MR Academy of Pediatrics surgical section survey—1973. J Pediatr Surg. imaging of fetal central nervous system abnormalities 1. Radiology. 1974;9(3):389–98. 2003;229(1):51–61. 67. Avni FE, Guibaud L, Robert Y, Segers V, Ziereisen F, Delaet MH, Metens T. 64. Griffiths PD, Paley MN, Whitby EH. MR imaging of the fetal brain MR imaging of fetal sacrococcygeal teratoma: diagnosis and assessment. and spine: a maturing technology. Ann Acad Med Singapore. Am J Roentgenol. 2002;178(1):179–83. 2003;32(4):483–9.

Submit your next manuscript to BioMed Central and we will help you at every step: • We accept pre-submission inquiries • Our selector tool helps you to ﬁnd the most relevant journal • We provide round the clock customer support • Convenient online submission • Thorough peer review • Inclusion in PubMed and all major indexing services • Maximum visibility for your research

Submit your manuscript at www.biomedcentral.com/submit